Starting point
AI models aren't human.
They don't have persistent memory, unless you tell them what to remember. They get "tired" when their context runs out. They don't communicate across sessions, unless you tell them to. Yet in many ways, humans aren't much better.
ClaudeKit is the designer's answer to an engineering problem. How can we help humans think more like the AIs they collaborate with? And how can we apply human accountability processes to AIs so the work they do is higher quality before human eyes see it?
What we found
Everyone can learn.
Both humans and AIs respond to feedback. We try, we fail, we adjust, we try again. What was missing in our AI development systems was the feedback loop: we didn't know how to get better, and neither did Claude. We both kept making the same mistakes when doing our parts of the work. As designers, we value feedback as the highest gift. So we built a tool to integrate it into every session, automatically.
Where we went
A mentor, a critic, and a scorecard.
The Mentor reads your prompt immediately, before sending it off to be executed. It evaluates how clear you have been with your intention and your directions. It gives you feedback about what you did well, and what will cost you, either in the AI making assumptions or taking more resources to hunt down answers itself. It even grades you and keeps track of your scores in a dashboard so you can try to improve over time. Each week it checks whether its own advice has gone stale, researches what has changed in Claude Code, and teaches you the new tricks.
The Critic responds when called. Whenever you want the AI to verify its work before showing you, you can say "show your progress to the Critic and take its feedback before showing me," or simply "critic" which does the same thing. This spawns a separate agent who is fed two things: your instructions and the work that was done. Then it critiques the work: did it execute the instructions? Lessons from the failures are fed into a document called Hard Won Lessons so that the AI can also learn from its mistakes.
What's next
More tools.
The Mentor, the Critic and the scorecard are the tools that made it out of our R&D phase, because they were the most genuinely useful. But ClaudeKit is ever-evolving, as both we and our AI pipelines develop.
Clone the repo and add it to your next Claude Code project. Try it out. And most importantly: give us feedback.